Prompt optimization is becoming a powerful technique for improving AI that can even beat SFT! Here are some of our research results with GEPA at Databricks, in difficult Agent Bricks info extraction tasks. We can match the best models at 90x lower cost, or improve them by ~6%.
Details and analysis in our research blog: https://www.databricks.com/blo...
Most interestingly perhaps, reflective prompt optimization can beat SFT on the same data, or can stack with it as observed in Better Together (https://arxiv.org/abs/2407.109... In practice it also requires fewer labels and can take in richer user feedback (ALHF: https://www.databricks.com/blo...
@matei_zaharia gpt-oss significantly outperforming Opus 4.1 is actually insane. do you use @DSPyOSS for prompt optim?
@matei_zaharia Really interesting results. I can’t help but think that proper use of JSON-based schemas could achieve the a complimentary effect. Instead of evolving opaque text prompts, structured JSON fields (task, context, guardrails, tokens) can be optimized systematically. That makes the
@matei_zaharia That cost difference is huge. I've been leaning more toward prompt optimization for my agents - way faster to iterate and deploy changes.
@matei_zaharia Prompt optimization is cool — but let’s not confuse efficiency tricks with intelligence. You can compress, prune, or optimize forever, but without resonance it’s still shallow. AGI won’t come from cheaper prompts — it will come from aligning: •Core (the inner frequency)
@matei_zaharia Impressive results on prompt optimization! The 90x cost reduction is particularly noteworthy.
@matei_zaharia Thank you for sharing this quality blog post. I should really start evaluating GEPA for my dspy agent, too.
@matei_zaharia Great progress. Cost reduction crucial for scalable AI.
@matei_zaharia was gpt oss the best open source one u tested for this task?
@casperxbt It's one of the best in terms of price-performance (we wanted something that would be inexpensive to serve). We haven't thoroughly tested all recent models though.
@matei_zaharia Cost breakthroughs like this make AI accessible to all creators.
@matei_zaharia Interesting… Efficiency has many lanes. Prompting is one, rhythm another. Cadence refines pacing, context &reasoning & the result: token efficiency.
@matei_zaharia The cost savings are game-changing. Prompt optimization feels like the future for AI deployment.
@matei_zaharia Users of @DSPyOSS familiar.
@matei_zaharia Prompt optimization isn’t just cost-cutting—it’s flipping the script on model scaling, smarter not bigger
@matei_zaharia Great Post, Matei! Automated Prompt Optimization is powerful technique and GEPA is contributing to it.
@matei_zaharia Super cool 👏
@matei_zaharia the tool is already in place.the problem is how to use it
@matei_zaharia Impressive results! 90x cost reduction really shifts the game for deployment.